MolE-RTD-ZINC1.5B-S2

Downstream-ready molecular encoder derived from caithmac/MolE-RTD-ZINC1.5B.

Step 1 used 25% replaced-token detection over approximately 1.538B ZINC20 molecular presentations. Step 2 used ChEMBL 36 sparse multitask supervision over 642 assays.

Recommended checkpoint

encoder_weights.pt is the fixed 50,000-step Step-2 snapshot used for the reported 22-task TDC evaluation.

SHA-256: 7940ca66fc8f6785d6ae2e63b3965f69425e872cd28f4a975493fb1972742ee1

All exported 10K-80K Step-2 snapshots are in checkpoints/. The 50K snapshot was fixed before test evaluation; it was not selected using TDC test scores. training_state/training_state_latest.pt is the resumable 80K training state, not an inference checkpoint. Evaluation summaries and provenance are in results/.

Loading

import torch
from huggingface_hub import hf_hub_download
from encoder_arch import resolve_encoder_config
from DeBERTa.deberta.config import ModelConfig
from mole.training.models.mole import AtomEnvEmbeddings

path = hf_hub_download(
    "caithmac/MolE-RTD-ZINC1.5B-S2",
    "encoder_weights.pt",
)
state = torch.load(path, map_location="cpu", weights_only=False)
config = resolve_encoder_config(state, requested="rtd25_step1")

encoder = AtomEnvEmbeddings(ModelConfig.from_dict(config))
encoder.load_state_dict(state, strict=True)
encoder.eval()

See FINETUNE_AND_REPRESENT.md for downstream examples.

Based on Recursion Pharmaceuticals' MolE and Microsoft DeBERTa-v3. Released under the inherited CC BY-NC 4.0 terms.